The Miracle of Microfinance? Evidence from a
Randomized Evaluation
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia
Kinnan. “The Miracle of Microfinance? Evidence from a
Randomized Evaluation.” American Economic Journal: Applied
Economics 7, no. 1 (January 2015): 22–53. © American
Economic Association
As Published
http://dx.doi.org/10.1257/app.20130533
Publisher
American Economic Association
Version
Final published version
Accessed
Thu May 26 03:07:20 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/95941
Terms of Use
Article is made available in accordance with the publisher's policy
and may be subject to US copyright law. Please refer to the
publisher's site for terms of use.
Detailed Terms
American Economic Journal: Applied Economics 2015, 7(1): 22–53
http://dx.doi.org/10.1257/app.20130533
The Miracle of Microfinance?
Evidence from a Randomized Evaluation †
By Abhijit Banerjee, Esther Duflo, Rachel Glennerster,
and Cynthia Kinnan *
This paper reports results from the randomized evaluation of a
group-lending microcredit program in Hyderabad, India. A lender
worked in 52 randomly selected neighborhoods, leading to an
8.4 percentage point increase in takeup of microcredit. Small business investment and profits of preexisting businesses increased, but
consumption did not significantly increase. Durable goods expenditure increased, while “temptation goods” expenditure declined.
We found no significant changes in health, education, or women’s
empowerment. Two years later, after control areas had gained access
to microcredit but households in treatment area had borrowed for
longer and in larger amounts, very few significant differences persist.
(JEL G21, G31, O16, O12, L25, I38)
M
icrofinance institutions (MFIs) have expanded rapidly over the last 10
to 15 years: according to the Microcredit Summit (Microcredit Summit
Campaign 2012), the number of very poor families with a microloan has grown
more than 18-fold from 7.6 million in 1997 to 137.5 million in 2010. Microcredit
has generated considerable enthusiasm and hope for fast poverty alleviation, culminating in the Nobel Prize for Peace, awarded in 2006 to Mohammed Yunus and
the Grameen Bank for their contribution to the reduction in world poverty. In the
last several years, h­ owever, the enthusiasm for microcredit has been matched by
* Banerjee: MIT Department of Economics, 40 Ames Street E17-201A, Cambridge, MA 02142 and National Bureau
of Economic Research (NBER) and J-PAL (e-mail: banerjee@mit.edu); Duflo: MIT Department of Economics, 40
Ames Street E17-201B, Cambridge, MA 02142 and NBER and J-PAL (e-mail: eduflo@mit.edu); Glennerster: J-PAL.
30 Wadsworth Street E53-320, Cambridge, MA 02142 (e-mail: rglenner@mit.edu); Kinnan: Northwestern University
Department of Economics, 2001 Sheridan Road 3222, Evanston, IL 60208 and NBER and J-PAL (e-mail: c-kinnan@
northwestern.edu). This paper updates and supersedes the 2010 version, which reported results using one wave of endline surveys. Funding for the first wave of the survey was provided by The Vanguard Charitable Endowment Program
and ICICI bank. Funding for the second wave was provided by Spandana and J-PAL. This draft was not reviewed by The
Vanguard Charitable Endowment Program, ICICI Bank, or Spandana. The Centre for Micro Finance at the Institute for
Financial Management Research (IFMR) (Chennai, India) set up and organized the experiment and the data collection,
and made the anonymized data available first to the research team, and then publicly. At the time, IFMR did not have
an IRB. Data analysis and on-going data collection have received IRB approval from MIT COUHES (1203004973)
and Northwestern University (STU00063636). Aparna Dasika and Angela Ambroz provided excellent assistance in
Hyderabad. Adie Angrist, Leonardo Elias, Harris Eppsteiner, Shehla Imran, Seema Kacker, Tracy Li, Aditi Nagaraj, and
Cecilia Peluffo provided excellent research assistance. Datasets for both waves of data used in this paper are available
at http://www.centre-for-microfinance.org/publications/data/. The authors wish to extend thanks to CMF and Spandana
for organizing the experiment, to Padmaja Reddy CEO of Spandana whose commitment to understanding the impact
of microfinance made this project possible, to Annie Duflo (the executive director of CMF at the time of the study) for
setting up this project, and to numerous seminar audiences and colleagues for insightful suggestions.
† Go to http://dx.doi.org/10.1257/app.20130533 to visit the article page for additional materials and author
disclosure statements or to comment in the online discussion forum.
Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
23
an equally strong backlash. For instance, a November 2010 article in the New York
Times, appearing in the wake of a rash of reported suicides linked to over-indebtedness, quotes Reddy Subrahmanyam, an official in Andhra Pradesh (the setting of
this study), accusing MFIs of making “hyperprofits off the poor.” He argues that
“the industry [has] become no better than the widely despised village loan sharks it
was intended to replace.... The money lender lives in the community. At least you
can burn down his house. With these companies, it is loot and scoot” (Polgreen and
Bajaj 2010).
What is striking about this debate is the relative paucity of evidence to inform it.
Anecdotes about highly successful entrepreneurs or deeply indebted borrowers tell
us nothing about the effect of microfinance on the average borrower, much less the
effect of having access to it on the average household. Even representative data about
microfinance clients and nonclients cannot identify the causal effect of microfinance
access, because clients are self-selected and therefore not comparable to nonclients.
Microfinance organizations also purposely choose some villages and not others in
which to operate. These issues make the evaluation of microcredit particularly difficult, and until recently there was little rigorous evidence on the impact of microfinance.
This has changed in the last few years, as several studies evaluating microfinance
have been conducted by different research teams with different partners in different
settings: Morocco (Crépon et al. 2015), Bosnia-Herzegovina (Augsburg et al. 2015),
Mexico (Angelucci, Karlan and Zinman 2015), Mongolia (Attanasio et al. 2015), and
Ethiopia (Tarozzi, Desai and Johnson 2015). In this paper we report on the oldest
of these, the first randomized evaluation of the effect of the canonical group-lending microcredit model, which targets women who may not necessarily be entrepreneurs. This study also follows the households over the longest period of any evaluation
(3 to 3.5 years after the introduction of the program in their areas), which is necessary
since many impacts may be expected to surface only over the medium run.
The experiment, a collaborative project between the Centre for Micro Finance
(CMF) at the Institute for Financial Management Research (IFMR) in Chennai and
Spandana, one of India’s fastest growing MFIs at the time, was conducted as follows. In 2005, 52 of 104 poor neighborhoods in Hyderabad were randomly selected
for the opening of a Spandana branch, while the remainder were not.1 Hyderabad
is the fifth largest city in India, and the capital of Andhra Pradesh, the Indian state
where microcredit has expanded the fastest and where it has been most controversial in recent years. Fifteen to 18 months after the introduction of microfinance in
each area, a comprehensive household survey was conducted with an average of
65 households in each neighborhood, for a total of about 6,850 households. In the
meantime, other MFIs had also started their operations in both treatment and comparison areas, but the probability of receiving an MFI loan was still 8.4 percentage
1 An alternative way to measure the impact of borrowing is to randomize microcredit offers among applicants.
This approach was pioneered by Karlan and Zinman (2010), who use individual randomization of the “marginal”
clients in a credit scoring model to evaluate the impact of consumer lending in South Africa, finding that access
to microcredit increases the probability of employment. The authors use the same approach to measure impact of
microcredit among small businesses in Manila (Karlan and Zinman 2011). It should be noted, however, that these
two studies evaluate slightly different programs: consumer lending in the South Africa study, and “second generation” individual-liability loans to existing entrepreneurs in Manila. 24
American Economic Journal: Applied Economics
January 2015
points (46 percent) higher in treatment areas than in comparison areas (26.7 percent
borrowers in treated areas versus 18.3 percent borrowers in comparison areas). Two
years after this first endline survey, the same households were surveyed once more.
By that time, both Spandana and other organizations had started lending in the treatment and control groups, so the fraction of households borrowing from microcredit
organizations was not dramatically different (38.5 percent in treatment and 33 percent in control). But households in treatment groups had larger loans and had been
borrowing for a longer time period. This second survey thus gives us an opportunity
to examine some of the longer term impacts of microcredit access on households
and businesses, although the setting is not perfect since we are comparing those who
borrow for longer versus those who borrow for a shorter time, rather than those who
borrow and those who do not borrow at all.
Since it is entirely possible that there are spillover or general equilibrium effects
(as analyzed by Buera, Kaboski, and Shin 2011), and effects that operate through
the expectation of being able to borrow when needed (such as reductions in precautionary savings, as documented in Thailand by Kaboski and Townsend (2011) and in
India by Fulford (2011)), or through general equilibrium effects on prices or wages
(Giné and Townsend 2004), we focus here on reduced-form/intent-to-treat estimates.
We examine the effect on borrowing from various sources, consumption, new
business creation, business income, etc., as well as on measures of other human
development outcomes, such as education, health, and women’s empowerment. At
the first endline, while households do borrow more from microcredit institutions,
the overall take up is reasonably low (only 26.7 percent of the eligible households
borrow, not the 80 percent that Spandana expected) and some microloans are substitutes for informal loans. Informal borrowing declines, and we see no significant
difference in overall borrowed amount (though the point estimate is positive). This
in itself was a surprising result at the time, though it has been replicated in other
studies: the demand for microcredit is less than expected, and may not correspond
to an important demand for additional credit. We see no significant difference in
monthly per capita consumption or monthly nondurable consumption. We do see
significant positive impacts on the purchase of durables. There is evidence that this
is financed partly by an increase in labor supply and partly by cutting unnecessary consumption: households have reduced expenditures on what they themselves
describe as “temptation goods.”
Thus, in our context, microfinance plays a role in helping some households
make different intertemporal choices in consumption. This is not the only impact
that is traditionally expected from microfinance, however. The primary engine of
growth that it is supposed to fuel is business creation.2 This is typically true even
for lenders that do not insist that households have a business to take a first loan
(Spandana is one of them), but still hope and expect that the ability to borrow will
eventually help households start or expand small businesses. (The description of
Spandana’s group-loan product is careful not to mention an automatic link between
2 To give a sense of the prevalence of the purported link between microfinance and business creation, of the
roughly 3.1 million Google search results for “microfinance,” 1.35 million (44 percent) also contain the phrase
“business creation” or “entrepreneurship” (retrieved November 2013). Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
25
credit and s­elf-­employment activity, but does state that “Loans are used for cash
flow ­smoothening [sic], predominantly for productive purposes.”) Fifteen to 18
months after gaining access, households are no more likely to be entrepreneurs (that
is, have at least one business), but they invest more in the businesses they do have
(or the ones they start). There is an increase in the average profits of the businesses
that were already in existence before microcredit, which is entirely due to very large
increases in the upper tail of profitability. At every quantile between the fifth and
the ninty-fifth percentile, there is no difference in the profits of the businesses. The
median marginal new business is both less profitable and less likely to have even one
employee in treatment than in control areas.
After three years, when microcredit is available both in treatment and control
groups but treatment group households have had the opportunity to borrow for a
longer time, businesses in the treatment groups have significantly more assets, and
business profits are now larger for businesses above the eighty-fifth percentile of
profitability. However, the average business is still small and not very profitable.
In other words, perhaps contrary to most people’s belief, to the extent microcredit
helps businesses, it may help the most profitable businesses the most. There is still
no difference in average consumption.
We do not find any effect on any of the women’s empowerment or human development outcomes we examine, either after 18 or 36 months. Furthermore, almost
70 percent of eligible households do not have an MFI loan, preferring instead to
borrow from other sources, if they borrow (and most do).
A number of caveats must be kept in mind when interpreting and generalizing these
results. First, the difference in microfinance take-up between treatment and control
areas is low, even at the first endline, which raises two issues: it lowers power and precision (though we have a number of significant effects), and it means that the impact
of microcredit we detect is driven by marginal borrowers—those who do not borrow
when the cost of doing so is high (because they have fewer MFIs to choose from or do
not want to change neighborhoods), but do borrow when that cost is lower.
Second, the evaluation was run in a context of very high economic growth, which
could have either decreased or increased the impact of microfinance. Third, this is
the evaluation of a for-profit microfinance model; not-for-profit microfinance lenders may have larger positive effects if their interest rates are kept low. Fourth, as
the MFI we study does not provide any complementary services, such as business
training or sensitivity education, we are studying the pure impact of providing loans
to women who may or may not use them for their own businesses (though Spandana
does believe that this is what the money will be used for eventually, and we do find
an expansion in women-owned businesses). Fifth, the study took place in “marginal” neighborhoods—those Spandana was indifferent about working with at the
outset—and the impacts may have been different in the neighborhoods they chose
to exclude from the randomization (Heckman 1992).
Thus, it is an important reassurance that our results find a strong echo in five other
studies that look at similar programs in different contexts (discussed below). This
gives us confidence in the robustness and external validity of our findings. In short,
microcredit is not for every household, or even most households, and it does not lead
to the miraculous social transformation some proponents have claimed. Its p­ rincipal
26
American Economic Journal: Applied Economics
January 2015
impact seems to be, perhaps unsurprisingly, that it allows some households to sacrifice some instantaneous utility (temptation goods or leisure) in order to finance
lumpy purchases, either for their home or in order to establish or expand a business. Prima facie, these marginal businesses do not appear to be highly p­ roductive
or profitable, but more data and more time may be needed to fully establish their
impacts on individuals, markets, and communities.
I. The Spandana Microcredit Product and the Context
A. Spandana and Its Microcredit Product
Until the major crisis in Indian microfinance in 2010, Spandana was one of the
largest and fastest growing microfinance organizations in India, with 1.2 million
active borrowers in March 2008, up from 520 borrowers in 1998, its first year of
operation (MIX Market 2009). It had expanded from its birthplace in Guntur, a
dynamic city in Andhra Pradesh, across the state and into several others.
The basic Spandana product is the canonical group-loan product, first introduced
by the Grameen Bank. A group is comprised of 6 to 10 women, and 25–45 groups
form a “center.” Women are jointly responsible for the loans of their group. The
first loan is Rs. 10,000, about $200 at market exchange rates, or $1,000 at 2007
purchasing power parity (PPP)-adjusted exchange rates (World Bank 2007).3 It
takes 50 weeks to reimburse principal and interest; the interest rate is 12 percent
(­nondeclining balance; equivalent to a 24 percent APR). If all members of a group
repay their loans, they are eligible for second loans of Rs. 10,000–12,000. Loan
amounts increase up to Rs. 20,000. During the course of the study, Spandana also
introduced an individual product, for clients who had been successful with one or
two group-loan cycles. The individual product was available in the treatment areas.
Very few people in our sample ended up taking this loan, however, so the study is
mainly an evaluation of a group-lending product.
Eligibility is determined using the following criteria. Clients must (i) be female,
(ii) be aged 18 to 59, (iii) have resided in the same area for at least one year, (iv)
have valid identification and residential proof (ration card, voter card, or electricity
bill), and (v) at least 80 percent of women in a group must own their home.4 Groups
are formed by women themselves, not by Spandana.
Unlike some other microfinance organizations, Spandana does not require its clients to start a business (or pretend to) in order to borrow: the organization recognizes
that money is fungible, and clients are left entirely free to choose the best use of the
money, as long as they repay their loan. Spandana does not determine loan eligibility by the expected productivity of the investment, although selection into groups
may screen out women who cannot convince fellow group members that they are
likely to repay. Also, unlike other microlenders (most notably Grameen) Spandana
3 In 2007 the PPP exchange rate was $1 = Rs. 9.2, while the market exchange rate was $1 ≃
​ ​Rs. 50. All following references to dollar amounts are in PPP terms unless noted otherwise. 4 The home ownership requirement is not because the house is used as collateral, but because home owners are
more stable and less likely to migrate. Spandana does not require a formal property title, just a general agreement
that this house belongs to this household (something that tends to be clear even in informal settlements). Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
27
does not explicitly insist on “transformation” in the household. There is no chanting
of resolutions at group meetings, which are very short and focused on the repayment
transaction. Spandana is primarily a lending organization, not directly involved in
business training, financial literacy promotion, etc. It is, however, the belief of the
management that the very fact of borrowing will lead to such transformation and
to business creation. Spandana is also a for-profit operator, charging interest rates
sufficient to make profits, though all the profits were re-invested in the organization
in the period we study. The organization obtained private capital and would probably have launched an IPO if it had not been caught in the middle of the Andhra
Pradesh microfinance crisis in 2010. This makes it different from Grameen Bank
(Mohammed Yunus has explicitly and vigorously criticized for-profit MFIs after
the IPO of Compartamos, a large Mexican MFI). All these features are important
to keep in mind when interpreting the results of this study; it is possible that a
Grameen-type organization would have different impacts. However, from an evaluation point of view, there are clear advantages to this product: in particular, any
impact on business expansion, etc. can be attributed to credit alone, rather than to
other services. Moreover, to the extent we find “positive” results in the study, they
are unlikely to be attributable to social desirability bias. It is also worth noting that,
in the period we study, the interest rates charged by Spandana were low by typical
microfinance standards, even when compared to rates charged by Grameen.
B. The Context
Table 1A uses the baseline data to show a snapshot of households from the study
area in 2005, before the Spandana product was launched. As we describe below, these
numbers need to be viewed with some caution, as the households sampled at baseline
were not necessarily representative of the area as a whole, and were not purposely resurveyed at endline. At baseline, the average household was a family of five, with monthly
expenditure of just under Rs. 5,000, or $540 at PPP-adjusted exchange rates ($108
per capita) (World Bank 2005).5 There was almost no MFI borrowing in the sample
areas at baseline. However, 68 percent of the households had at least one outstanding
loan. The average amount outstanding was Rs. 38,000. Sixty-three percent of house­
holds had a loan from an informal source (moneylenders, friends or neighbors, family
members, or shopkeepers). Commercial bank loans were very rare (3.6 percent).
Although business investment was not commonly named as a motive for borrowing, businesses were common, with 32 businesses per 100 households, compared to
an OECD-country average of 12 percent who say that they are self-employed. Less
than half of all businesses were operated by women (14.5 woman-run businesses per
100 households.) Business owners and their families spent on average 76 hours per
week working in the business.
Growth between 2005 and 2010.—Table 1B shows some of the same key statistics for the endline 1 and endline 2 (EL1 and EL2) samples in the control group.
5 Column 2 reports the control mean and column 4 reports the treatment-control difference. Only one difference
out of 33 is significant at the 10 percent level (column 5). 28
American Economic Journal: Applied Economics
January 2015
Table 1A—Baseline Summary Statistics
Control group
Mean
(2)
Household composition
Number members
Number adults (>=16 years old)
Number children (<16 years old)
Male head
Head’s age
Head with no education
1,220
1,220
1,220
1,216
1,216
1,216
5.038
3.439
1.599
0.907
41.150
0.370
(1.666)
(1.466)
(1.228)
(0.290)
(10.839)
(0.483)
0.095
−0.011
0.104
−0.012
−0.243
−0.008
0.303
0.873
0.098
0.381
0.676
0.787
Access to credit
Loan from Spandana
Loan from other MFI
Loan from a bank
Informal loan
Any type of loan
1,213
1,213
1,213
1,213
1,213
0.000
0.011
0.036
0.632
0.680
(0.000)
(0.103)
(0.187)
(0.482)
(0.467)
0.007
0.007
0.001
0.002
0.002
0.195
0.453
0.859
0.958
0.942
Amount borrowed from (in Rs)
Spandana
Other MFI
Bank
Informal loan
Total
1,213
1,213
1,213
1,213
1,213
0
201
7,438
28,460
37,892
(0.000)
(2,742)
(173,268)
(65,312)
(191,292)
69
170
−5,420
−570
−5,879
0.192
0.568
0.279
0.856
0.343
Self-employment activities
Number of activities
Number of activities managed by women
Share of HH activities managed by women
1,220
1,220
295
0.320
0.145
0.488
(0.682)
(0.400)
(0.482)
−0.019
−0.007
−0.006
0.579
0.750
0.904
Businesses
Revenue/month (Rs)
Expenses/month (Rs)
Investment/month (Rs)
Employment (employees)
Self-employment (hours per week)
295
295
295
295
295
15,991
3,617
385
0.169
76.315
(53,489)
(26,144)
(3,157)
(0.828)
(66.054)
4,501
641
14
0.255
−4.587
0.539
0.751
0.959
0.148
0.414
Businesses (all households)
Revenue/month (Rs)
Expenses/month (Rs)
Investment/month (Rs)
Employment (employees)
Self-employment (hours per week)
1,220
1,220
1,220
1,220
1,220
3,867
875
93
0.041
18.453
(27,147)
(12,933)
(1,559)
(0.413)
(46.054)
904
116
−0.098
0.057
−1.801
0.626
0.812
0.999
0.166
0.400
1,220
1,220
1,220
1,220
4,888
4,735
154
1.941
(4,074)
(3,840)
(585)
(0.829)
270
252
18
0.027
0.232
0.235
0.531
0.669
Consumption ( per household per month)
Total consumption (Rs)
Nondurables consumption (Rs)
Durables consumption (Rs)
Asset index
SD
(3)
Treatment − control
Obs.
(1)
Coeff.
(4)
p-value
(5)
Notes: Unit of observation: household. Standard errors of differences, clustered at the area level, in parentheses.
Sample includes all households surveyed at baseline. Informal lender includes moneylenders, loans from friends/
family, and buying goods/services on credit from seller. Asset index is calculated on a list of 40 home durable
goods. Each asset is given a weight using the coefficients of the first factor of a principal component analysis. The
index, for a household i, is calculated as the weighted sum of standardized dummies equal to 1 if the household
owns the durable good.
Source: Baseline household survey
Banerjee et al.: The Miracle of Microfinance?
Vol.7 No. 1
29
Table 1B—Endline 1 and 2 Summary Statistics (Control group)
EL1 control group
EL2 control group
EL2-EL1
Obs.
(1)
Mean
(2)
SD
(3)
Obs.
(4)
Mean
(5)
SD
(6)
Coeff.
(7)
p-value
(8)
3,264
3,264
5.645
3.887
(2.152)
(1.754)
2,943
2,943
6.269
4.039
(2.548)
(1.848)
0.624
0.152
0.000
0.000
Household composition
Number members
Number adults
(>=16 years old)
Number children
(<16 years old)
Male head
Head’s age
Head with no education
3,264
1.738
(1.310)
2,943
1.764
(1.321)
0.026
0.247
3,261
3,257
3,256
0.895
41.149
0.311
(0.307)
(10.222)
(0.463)
2,938
2,940
2,940
0.811
42.258
0.292
(0.391)
(10.154)
(0.455)
−0.083
1.109
−0.020
0.000
0.000
0.021
Access to credit
Loan from Spandana
Loan from other MFI
Loan from a bank
Informal loan
Any type of loan
3,247
3,183
3,247
3,247
3,264
0.051
0.149
0.079
0.761
0.867
(0.219)
(0.356)
(0.270)
(0.427)
(0.339)
2,943
2,943
2,943
2,943
2,943
0.111
0.268
0.073
0.603
0.904
(0.315)
(0.443)
(0.260)
(0.489)
(0.294)
0.061
0.120
−0.006
−0.158
0.037
0.000
0.000
0.480
0.000
0.000
Amount borrowed from (in Rs)
Spandana
Other MFI
Bank
Informal loan
Total
3,247
3,200
3,247
3,247
3,264
597
1,806
8,422
41,045
59,836
(2,907)
(5,918)
(101,953)
(78,033)
(133,693)
2,943
2,943
2,943
2,943
2,943
1,567
4,775
6,127
32,356
88,631
(5,618)
(10,736)
(40,308)
(76,704)
(144,634)
969
2,969
−2,296
−8,689
28,795
0.000
0.000
0.221
0.000
0.000
3,236
3,209
0.503
0.185
(0.854)
(0.487)
2,943
2,943
0.561
0.234
(0.787)
(0.520)
0.058
0.050
0.003
0.000
1,104
0.377
(0.453)
1,231
0.403
(0.454)
0.026
0.113
Businesses
Revenue/month (Rs)
Expenses/month (Rs)
Investment/month (Rs)
Employment (employees)
Self-employment (hrs/wk)
1,039
1,071
1,127
1,103
1,103
14,700
12,030
785
0.380
100.03
(56,350)
(51,531)
(6,806)
(1.644)
(69.87)
1,218
1,218
1,231
1,231
1,231
14,066
12,568
2,331
0.565
88.47
(23,713)
(30,483)
(14,645)
(2.938)
(60.16)
−634
538
1,546
0.185
−11.56
0.724
0.769
0.001
0.062
0.000
Businesses (all households)
Revenue/month (Rs)
Expenses/month (Rs)
Investment/month (Rs)
Employment (employees)
Self-employment (hrs/wk)
3,145
3,177
3,231
3,209
3,209
4,856
4,055
280
0.131
34.13
(33,108)
(30,446)
(4,038)
(0.980)
(62.59)
2,930
2,930
2,943
2,943
2,943
5,847
5,225
1,007
0.236
37.00
(16,784)
(20,603)
(9,623)
(1.920)
(58.46)
991
1,169
727
0.106
2.88
0.105
0.088
0.001
0.011
0.037
(4,906)
(4,212)
(1,623)
(0.861)
2,943
2,943
2,941
2,943
8,787
8,050
720
2.662
(6,547)
(5,780)
(1,536)
(0.828)
2,412
2,219
169
0.291
0.000
0.000
0.000
0.000
Self-employment activities
Number of activities
Number activities
managed by women
Share activities
managed by women
Consumption ( per household per month)
Consumption
3,248
6,375
Nondurables consumption
3,230
5,831
Durables consumption
3,230
551
Asset index
3,254
2.371
Notes: Summary statistics for comparison areas only. Standard errors of differences, clustered at the area level, in parentheses (column 3). All monetary amounts in 2007 Rs. Asset index is calculated on a list of 40 home durable goods. Each
asset is given a weight using the coefficients of the first factor of a principal component analysis. The index, for a household i, is calculated as the weighted sum of standardized dummies equal to 1 if the household owns the durable good.
Comparing the control baseline sample (2005) with the control households in the
EL1 (2008) and EL2 (2010) samples reveal very rapid secular growth in Hyderabad
over 2005–2010.6 Average household consumption rose from Rs. 4,888 (2005) to
6 While the comparison may not be perfect since the baseline survey was not conducted on the same sample as
the endline, the growth between EL1 and EL2 is for the same set of households, using the same survey instruments,
and thus gives us a good sense of the dynamism of this economy. 30
American Economic Journal: Applied Economics
January 2015
Rs. 6,375 in 2007 and Rs. 8,787 in 2010 (all expressed in 2007 rupees). The fraction
of households with at least one outstanding loan rose from 68 percent at baseline to
87 percent in EL1 and 90 percent in EL2.
The prevalence of businesses increased from 32 per hundred households at baseline
to 50 at EL1 and 56 at EL2. At endline 1, 37.7 percent, and at endline 2, 40.3 percent
of the businesses were operated by women. However, the businesses remained very
small, with, on average, 0.38 employees in EL1 and 0.57 in EL2. As well as remaining
very small in terms of employment, average sales remained fairly steady: Rs. 14,700
at EL1 and 14,100 at EL2. However, looking across all households (not just those
with businesses), business revenues increased from around Rs. 4,900 to Rs. 5,800 (in
constant 2007 rupees). At EL2, business owners reported business expenses (working
capital) plus investment in assets of almost Rs. 15,000, up from about Rs. 13,000 at
EL1. (These expense estimates do not account for the cost of the proprietors’ time.)
This context of rapid growth in urban Andhra Pradesh is another important feature
to keep in mind, and may color the results of this study; of all the randomized evaluations on microfinance, ours has probably the most dynamic context. The setting of this
study is clearly an important one, since microfinance clients in India represent roughly
30 percent of all microfinance clients worldwide,7 and since microfinance has developed in many other rapidly growing environments (Bangladesh being probably the
prime example). Nonetheless, the results of other evaluations of microfinance may be
different in contexts either with much slower growth or in recession. Fortunately, the
other five RCT studies of microfinance in this issue cover a wide variety of settings,
which will help to understand the extent to which results depend on context.
II. Experimental Design
A. Experimental Design
At the time this study was started, microfinance had already taken hold in several
districts in Andhra Pradesh, but most microfinance organizations had not yet started
working in the capital, Hyderabad. Spandana initially selected 120 areas (identifiable neighborhoods, or bastis) in Hyderabad as places in which they were interested
in opening branches but also willing not to do so. These areas were selected based
on having no preexisting microfinance presence and on having residents who were
desirable potential borrowers: poor, but not “the poorest of the poor.” Areas with high
concentrations of construction workers were avoided because they move frequently,
which makes them undesirable as microfinance clients. While the selected areas are
commonly referred to as “slums,” these are permanent settlements with concrete
houses and some public amenities (electricity, water, etc.). Conversely, the largest
such areas in Hyderabad were not selected for the study, since Spandana was keen to
start operations there: the large population in these slums allowed them to benefit from
economies of scale and quickly reach a number of clients that justified expansion in
the city. The population in the ­neighborhoods selected for the study ranges from 46 to
7 MIX Market reported 94 million borrowers worldwide in 2011, of whom 28 million were located in India
(http://www.mixmarket.org/mfi/country/India). Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
31
555 households. The slums chosen to be part of the study were typically not continuous to avoid spillovers across treatment and control slums.
In each area, CMF first hired a market research company to conduct a small
baseline neighborhood survey in 2005, collecting information on household composition, education, employment, asset ownership, expenditure, borrowing, saving, and any businesses currently operated by the household or stopped within
the last year. They surveyed a total of 2,800 households in order to obtain a rapid
assessment of the baseline conditions of the neighborhoods. However, since there
was no existing census, and the baseline survey had to be conducted very rapidly
to gather some information necessary for stratification before Spandana began
their operations, the households were not selected randomly from a household list:
instead, field officers were asked to map the area and select every n​ th​house, with​
n​ chosen to select 20 households per area. Unfortunately, this procedure was not
followed very rigorously by the market research company, and we are not confident that the baseline is representative of the slum as a whole. Thus, the baseline
survey was used solely as a basis for stratification, the descriptive analysis above,
and to collect area-level characteristics that are used as control variables.8 Beyond
this, we do not use the baseline survey in the analysis that follows.
After the baseline survey, but prior to randomization, 16 areas were dropped
from the study because they were found to contain large numbers of migrantworker households. Spandana (like other MFIs) has a rule that loans should only
be made to households who have lived in the same community for at least one
year because the organization believes that dynamic incentives (the promise of
more credit in the future) are more important in motivating repayment for these
households.9 The remaining 104 areas were grouped into pairs of similar neighborhoods, based on average per capita consumption and per-household debt, and
one of each pair was randomly assigned to the treatment group.10 Figure 1 shows
a timeline of data collection and randomization.
Table 1A uses the baseline sample to show that treatment and comparison areas
did not differ in their baseline levels of demographic, financial, or entrepreneurship
characteristics in the baseline survey. This is not surprising, since the sample was
stratified according to per capita consumption and fraction of households with debt.
Spandana then progressively began operating in the 52 treatment areas between
2006 and 2007. The rollout happened at different dates in different slums. Note
that in the intervening periods, other MFIs also started their operations, both in
treatment and in comparison areas. We will show that there is still a significant difference between MFI borrowing in treatment and comparison groups. Spandana
credit officers also started lending in very few of the control slums, although this
8 However, omitting these controls makes no difference to the results. We can compare baseline characteristics in the 16 areas dropped to those in the 104 areas included in the randomization. The differences are consistent with Spandana’s rationale for dropping the omitted areas: household size
is smaller in these areas (due to migrant workers there without families or children); there is less business creation
(presumably because migrants are unlikely to start a business); and there is less credit outstanding (likely because
informal lenders are also reluctant to lend to these very mobile households). (Results available upon request.) 10 Pairs were formed to minimize the sum across pairs A, B (area A avg. loan balance area B avg. loan balance)2 + (area A per capita consumption area B per capita consumption)2. Within each pair one neighborhood was
randomly allocated into treatment. 9 32
American Economic Journal: Applied Economics
January 2015
Endline sample
frame selection
Jul. ’07
Baseline
Jan ’05–Feb. ’06
Census
Endline 1*
Feb. ’07–Jan ’07 Aug. ’07–Apr. ’08
Endline 2
Nov. ’09–Jun. ’10
Jan. ’05 Jul. ’05 Jan. ’06 Jul. ’06 Jan. ’07 Jul. ’07 Jan. ’08 Jul. ’08 Jan. ’09 Jul. ’09 Jan. ’10 Jul. ’10
Spandana moves
into treatment areas
Apr. ’06–Apr. ’07
Spandana begins to
move into control areas
May ’08
Andhra Pradesh
microfinance crisis begins
Oct. ’10
Figure 1. Timeline of Intervention and Data Collection
Note: No treatment area was surveyed for endline 1 until at least one year had elapsed from the start of Spandana
lending in that area.
was stopped relatively rapidly. Furthermore, there was no rule against borrowing
in another slum (if one could find a group to join), and some people did do so.
Overall, 5 percent of households in control slums were borrowing from Spandana
at the time of the first endline.
To create a proper sampling frame for the endline, CMF staff undertook a
comprehensive census of each area in early 2007, and included a question on
borrowing. The census revealed low rates of MFI borrowing even in treatment
areas, so the endline sampling frame consisted of households whose characteristics suggested high likelihood of having borrowed: those that had resided in the
area for at least 3 years and contained at least 1 woman aged 18 to 55. Spandana
borrowers identified in the census were oversampled because we believed that heterogeneity in treatment effects would introduce more variance in outcomes among
Spandana borrowers than among nonborrowers, and that oversampling borrowers
would therefore give higher power. The results presented below weight the observation to account for this oversampling so that the results are representative of the
population as a whole. Since the sampling frame at baseline was not sufficiently
rigorous, baseline households were not purposely resurveyed in the follow-up.
The first endline survey began in August 2007 and ended in April 2008, and the
rollout of the endline followed the rollout of the program. This first endline survey
was conducted at least 12 months after Spandana began disbursing loans within
a given area, and generally 15 to 18 months after (the survey followed the same
calendar in the control slums, in order to ensure comparability between treatment
and control). The overall sample size was 6,863 households.
Two years later, in 2009–2010, a second endline survey, following up on the
same households, was undertaken. It included the same set of questions as in
2007–2008 to ensure comparability. The re-contact rate was very high (90 percent). We discuss this attrition in more detail below.
Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
33
B. Potential Threats to Identification and Caveats on Interpretation
Attrition and Selective Migration.—Since we lack a rigorous baseline sample that
was systematically followed, a potential worry is that the sample that was surveyed
at endline may not be strictly comparable in treatment and control areas, if there was
differential attrition in treatment and in control groups. For example, people could
have moved into the area, or avoided moving out of the area, because Spandana
had started their operations there. This does not seem highly likely, given that if
someone really wanted to borrow, they had options to do so either from another MFI
(we will see that a fair number of people did) or even from Spandana, by going to
another neighborhood. The treatment only made it marginally easier to borrow (as
we will see in the next section). Nevertheless, in retrospect, it was a clear mistake
not to attempt to systematically re-survey at least a fraction of the baseline sample,
even though the baseline sampling frame was weak.
That said, we have a number of ways to assess the extent to which attrition is a
problem. First of all, in Appendix Table A1,11 we verify that the households surveyed at endlines 1 and 2 are similar in treatment and control groups, in terms of a
number of characteristics that are fixed over time (the p-value on the joint difference
of these characteristics across treatment arms is 0.983 at EL1 and 0.567 at EL2).
This is a first indication that we have a comparable sample at baseline and at endline,
even allowing for attrition.
Second, the sample at EL1 was drawn from a census that was conducted fairly
soon after the introduction of microcredit (on average less than a year). Moreover,
the sampling frame for EL1 was restricted to people who had lived in the area for
at least three years before the census. This means that no one in the survey had
migrated into the area because of Spandana: they were all residents of the area well
before Spandana moved into the area (the vast majority had been there for years).
This removes the most plausible channel for differential selection into the sample
in treatment and control groups. There remains the possibility that fewer people (or
different people) left the treatment areas between the launch of the product and the
census due to the option to borrow more easily, but in less than a year, the migration
rate out of Hyderabad is low, and given the ability to borrow if someone wants to, it
seems far-fetched that people would have been differentially likely to migrate out of
the slums based on the ability to become a Spandana client.
We next examine attrition between the census and the first endline, and between
the first and second endlines. There was some attrition between the census and
EL1, especially since, as is customary in these types of surveys, census surveyors
were given replacement lists in case they did not find the exact person they were
looking for. However, this attrition (roughly 25 percent) is almost exactly the
same in treatment and in control areas: 27.6 percent in treatment and 25.2 percent in control ( p-value of difference: 0.332; see online Appendix Table A2,
panel A). Moreover, the attrition is totally uncorrelated with the months elapsed
since Spandana entered the slum (Table A2, panel B), which is not what we would
11 All Appendix tables are available in the online Appendix. 34
American Economic Journal: Applied Economics
January 2015
expect if it were ­somehow related to the program (it would have had more time to
play out if Spandana had entered a longer time before). The only characteristics
that predict that someone is more likely to be found is that they are a Spandana
borrower (4.2 percentage points lower attrition; SE of 1.97 percentage points),
and living in a “non-pucca” (lower-quality) house (2.7 percentage points lower
attrition; SE of 1.4 percentage points). The most likely reason for the former is
that the Spandana officers helped the CMF field team to locate their clients. For
example, surveyors could attend weekly meetings to collect addresses and find
directions to people’s homes. The latter likely reflects greater mobility among
wealthier households. In all of the analysis that follows, we correct for this by
adjusting the sampling weights for the ratio between the probability to find a
non-Spandana borrower and the probability to find a Spandana borrower (0.948 in
endline 1, 0.914 in endline 2).
Online Appendix Table A3, panel A shows that the re-contact rate at endline 2
for households initially interviewed at endline 1 was very high (much higher
than in most randomized controlled trials in either the United States or developing countries). It was also similar in the treatment and the control group, at 89.9
percent and 90.2 percent, respectively (the p-value of the difference is 0.248).
Panel B shows average characteristics of the re-contacted versus attrited households. The samples do not differ significantly along most dimensions. However,
those who attrited had slightly higher per capita expenditure at endline 1, with a
Rs. 1,000 increase in expenditure associated with a 0.0099 increase in likelihood
of attrition (column 1: the standard error is 0.0032). Having a Spandana loan at
endline 1 was associated with 3.4 percentage points lower attrition (column 5:
the standard error is 1pp); having any MFI loan is associated with 2.8 percentage points lower attrition (column 6: the standard error is 0.8pp), driven by the
effect of Spandana loans. Again, the explanation for this is that the credit officers
helped the field team find the clients, if they had moved within their slum. Panel
C of Table A3 shows that between treatment and control, attrition was not differentially correlated with characteristics, with the exception of having an MFI
loan (column 6), an effect likely driven by loan officers assisting in re-contacting
survey respondents.
This data suggests that there is no evidence that migration or attrition patterns
were driven by the treatment, except through the mechanical effect that Spandana
credit officers helped surveyors locate their clients, which we correct for.
Nevertheless, to systematically address the concern that attrition may affect the
results, we have re-estimated all the regressions below with a correction for sample
selection inspired by DiNardo, Fortin, and Lemieux (1996), where we re-weight
the data using the inverse of the propensity to be observed at endline 2, so that the
distribution of observable characteristics (at endline 1) among households observed
at endline 2 resembles that in the entire endline 1 sample. We then apply the same
weights to endline 1 data (implicitly assuming a similar selection process between
the onset of microfinance and endline 1). The results, presented for key outcomes
in online Appendix Table A5, are very similar to what we present here. (Full results
available on request.) Note that this procedure only corrects for differential attrition
by observables, not by unobservable variables.
Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
35
Interpreting the Results.—The experimental design and the implementation raise
a number of issues worth keeping in mind in interpreting the results that follow.
First, given the sampling frame, ours will be an intent-to-treat (ITT) analysis on
a sample of “likely borrowers.” This is thus neither the effect on those who borrow
nor the average effect on the neighborhood. Rather, it is the average effect of easier
access to microfinance on those who are its primary targets. Second, microfinance
was available in both treatment and control areas, though access was easier in treatment areas. Microfinance take-up is indeed higher in treatment areas, which generates experimental variation, but the marginal clients may be different from the first
clients to borrow in an area. This also affects power: the initial power ­calculations
were performed when Spandana thought that 80 percent of eligible households
would become clients very rapidly after the launch. In fact, the data shows that the
proportion reached only 18 percent in 18 months (and stayed at just below 18 percent after two and a half years). This is low, and also gave other MFIs, which were
behind Spandana in terms of penetration in Hyderabad, time to catch up. Overall,
take-up of microfinance from any organization was only 33 percent by EL2. This
is an important result in its own right, and very surprising at the time, but it implies
that, with the benefits of hindsight, more areas would have been needed. This is not
something that could be addressed ex post. Fortunately, subsequent evaluations of
microfinance programs were able to do so, and find a very similar set of results (and
nonresults), suggesting that these outcomes are not the artifact of samples that are
too small or of a very nonrepresentative set of clients.
III. Results
To estimate the impact of microfinance becoming available in an area on likely
clients, we focus on intent-to-treat (ITT) estimates; that is, simple comparisons of
averages in treatment and comparison areas, averaged over borrowers and non-­
borrowers. We present ITT estimates of the effect of microfinance on businesses
operated by the household; for those who own businesses, we examine business
profits, revenue, business inputs, and the number of workers employed by the business. (The construction of these variables is described in online Appendix A.) Each
column of each table reports the results of a regression of the form
​​y​ ia​  =  α + β × Trea​t ​ia​ + ​X′​ a ​​ γ + ​ε​ ia​​ ,
where ​​y​ ia​​is an outcome for household ​i​in area ​a​, ​Trea​t​ ia​​is an indicator for living in
a treated area, and β
​ ​is the intent-to-treat effect. ​X​ ′a ​​ is a vector of control variables,
calculated as area-level baseline values: area population, total businesses, average
per capita expenditure, fraction of household heads who are literate, and fraction of
all adults who are literate. Standard errors are adjusted for clustering at the area level
and all regressions are weighted to correct for oversampling of Spandana borrowers
and for higher probability of tracking them. We also estimated two sets of regressions with different specifications: one with no controls whatsoever, and one controlling for strata used in randomization rather than for the average characteristics
in the control slums. The results (available on request) are qualitatively unchanged.
36
American Economic Journal: Applied Economics
January 2015
Controlling for strata somewhat increases the precision in this case, so some results
that are almost significant here become significant with strata controls (this is particularly true for the grouped outcomes).
In any study of this kind, where there are many possible outcomes and multiple
possible causal pathways, there is a danger of overinterpreting any single significant
result (or even of discerning a pattern of results when there is none). We take a number
of steps to avoid this problem. First, we report outcomes following the template that
all papers in this issue follow, ensuring no selection of outcomes based on what is significant or not. Second, for each table (which corresponds to a “family” of outcomes)
we report an index (à la Kling, Liebman, and Katz 2007) of all the outcomes in the
family taken together.12 Finally, for each of these index outcomes, we report both
the standard p-value and the p-value adjusted for multiple hypotheses testing across
all the indices. The adjusted p-values are calculated using the step-down procedure
of Hochberg (1988), which controls the family-wise error rate for all the indices.13
A. Borrowing from Spandana and other MFIs
Treatment communities were randomly selected to receive Spandana branches,
but other MFIs also started operating both in treatment and comparison areas. We
are interested in testing the impact of access to microcredit, not only of borrowing from Spandana. Table 2, panel A shows that, by the first endline, MFI borrowing was indeed higher in treatment than in control slums, although borrowing
from other MFIs offset part of the difference in Spandana borrowing. Households in
treatment areas are 12.7 percentage points more likely to report being Spandana borrowers: 17.8 percent versus 5.1 percent (Table 2 panel A, column 1). The difference
in the percentage of households saying that they borrow from any MFI is 8.4 points
(Table 2 panel A, column 3), so some households who ended up borrowing from
Spandana in treatment areas would have borrowed from another MFI in the absence
of the intervention. While the absolute level of total MFI borrowing is not very high,
it is about 50 percent higher in treatment than in comparison areas. Columns 1 and
3 show that treatment households also report significantly higher loan amounts from
MFIs (and from Spandana in particular) than comparison households. Averaged
over borrowers and nonborrowers, treatment households report Rs. 1,334 more borrowing from Spandana than do control households, and Rs. 1,286 more from all
MFIs (both significant at the 1 percent level).
While both the absolute take-up rate and the implicit “first stage” are relatively
small, this result appears similar to what was found in most other evaluations of the
impact of access to microfinance, despite the different contexts. In rural Morocco,
Crépon et al. (2015) find that the probability of having any loan from the MFI Al
Amana in areas that received access to it is 10 percentage points, whereas it is
essentially 0 in control; moreover, since no other MFI operated in their study area,
12 The variables are signed such that a positive treatment effect is a “good” outcome. They are then normalized
by subtracting the mean in the control group and dividing by the standard deviation in the control group. The index
is the simple average of the normalized variables. 13 See online Appendix A4 for details. Banerjee et al.: The Miracle of Microfinance?
Vol.7 No. 1
37
Table 2—Credit
Spandana
(1)
Panel A. Endline 1
Credit access
Treated area
Observations
Control mean
Hochberg-corrected
p-value
0.127*** −0.012
(0.020)
(0.024)
6,811
0.051
Loan amounts (in Rupees)
Treated area
1,334***
(230)
Observations
Control mean
Panel B. Endline 2
Credit access
Treated area
Observations
Control mean
Hochberg-corrected
p-value
6,811
597
6,657
0.149
−94
(336)
6,708
1,806
0.063*** −0.039
(0.019)
(0.026)
6,142
0.111
Loan amounts (in Rupees)
Treated area
979***
(287)
Observations
Control mean
Other
MFI
(2)
6,142
1,567
6,142
0.268
−217
(628)
6,142
4,775
Any
MFI
(3)
Other
bank
(4)
0.084*** 0.003
(0.027)
(0.012)
6,811
0.183
6,811
0.079
1,286***
(439)
Informal
(5)
Total
(6)
−0.052** −0.023
(0.021)
(0.014)
6,811
0.761
6,862
0.867
75
(2,163)
−1,069
(2,520)
2,856
(4,548)
0.002
(0.029)
0.001
(0.009)
0.002
(0.018)
0.000
(0.010)
799
(669)
−1,181
(1,086)
158
(2,940)
2,554
(6,156)
6,811
2374
6,142
0.331
6,142
5,544
6,811
8,422
6,142
0.073
6,142
6,127
6,811
41,045
6,142
0.603
6,142
32,356
Number of
cycles
Ever
borrowed
Index of
late on
from an
dependent
payment?
MFI
variables
(7)
(8)
(9)
−0.060**
−0.026
6,475
0.616
0.084**
(0.041)
0.106***
(0.0291)
0.085
(0.067)
0.0288
(0.0253)
6,811
0.330
6,862
0.000
0.000
6,862
59,836
6,142
0.904
0.007
(0.021)
6,142
0.598
5,926
0.724
6,142
0.000
0.256
6,142
88,632
Notes: The table presents the coefficient of a “treatment” dummy in a regression of each variable on treatment (with control variables listed in the text). Cluster-robust standard errors in parentheses. Results are weighted to account for oversampling of Spandana
borrowers. Columns 1–6 under “Credit access” report the probability of having at least one loan from the source listed. The corresponding columns under “Loan amounts” report the loan amount (zero for nonborrowers). “Informal lender” includes moneylenders, loans from friends/family, and buying goods/services on credit. Number of loan cycles from an MFI is the maximum number
of loan cycles borrowed with a single MFI, including the current loan (if any); number of cycles is zero for MFI never-borrowers.
All monetary amounts in 2007 Rs. Column 9 presents the coefficient of a “treatment” dummy in a regression on treatment of an
index of z-scores of the outcome variables in columns 1–8 (including both credit access and loan amounts) for each round following
Kling, Liebman, and Katz (2007). p-values for this regression are reported using Hochberg’s step-up method to control the FWER
across all index outcomes. See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
this represents the total increase in microfinance borrowing. In Mexico, Angelucci,
Karlan, and Zinman (2015) find an increase of 10 percentage points in the probability of borrowing from the MFI Compartamos in areas that got access to the lender,
relative to a base of 5 percentage points in control. In Ethiopia, Tarozzi, Desai, and
Johnson (2015) find a larger impact of microcredit introduction: 36 percent.
The fairly low take-up rate in these different contexts is in itself a striking
result, given the high levels of informal borrowing in these communities and the
purported benefits of microcredit over these alternative forms of borrowing. In all
cases, except when the randomization was among those who had already expressed
explicit ­interest in microcredit (Attanasio et al. 2015 and Augsburg et al. 2015), only
a minority of “likely borrowers” end up borrowing.
38
American Economic Journal: Applied Economics
January 2015
18,000
6,000
0
−6,000
−18,000
OLS
Quantile treatment effect
90% C.I
5
10
20
30
40
50
60
70
80
90 95
Percentile
Figure 2. Treatment Effect on Informal Borrowing (Endline 1)
Notes: Informal borrowing: borrowing from moneylenders, friends and family, and buying goods
on credit. Confidence intervals are cluster-bootstrapped at the neighborhood level. For quantiles
0.05 to 0.20, confidence intervals are not reported because the quantile does not vary sufficiently
across neighborhoods to bootstrap standard errors. The point estimates are zero for these quantiles.
Table 2 also displays the impact of microfinance access on other forms of borrowing. A sizable fraction of the clients report repaying a more expensive debt as a
reason to borrow from Spandana, and we do indeed see some action on this margin.
The share of households who have some informal borrowing—defined as borrowing
from family, friends, or moneylenders or purchasing goods on credit extended by
the seller—goes down by 5.2 percentage points in treatment areas (column 5), but
bank borrowing is unaffected (column 4). The point estimate of the amount borrowed from informal sources is also negative, suggesting substitution of expensive
borrowing with cheaper MFI borrowing (an explicit objective of Spandana), and
the point estimate, though insignificant, is quite similar in absolute value to the
increase in MFI borrowing (column 3). However, given the high level of informal
borrowing, this corresponds to a decline of only 2.6 percent. When we examine the
distribution of endline 1 informal borrowing, in Figure 2, informal borrowing is significantly lower in treatment areas from the thirtieth to sixtieth percentiles. Overall,
treatment affects the index of borrowing outcomes, and the p-value is small even
when accounting for multiple hypothesis testing across families (column 9).
After the end of the first endline, following our initial agreement with Spandana,
Spandana started to expand into control areas. Other MFIs also continued their expansion. However, two years later, a significant difference still remained between treatment and control slums: Table 2, panel B shows that 17 percent of the households in
the treatment slums borrowed from Spandana, against 11 percent in the control slums.
Other MFIs continued to expand both in the former treatment and control slums, and
MFI lending overall was almost the same in the treatment and the control group. By the
second endline survey, 33.1 percent of households had borrowed from an MFI in the
former control slums, and 33.3 percent in the treatment slums. Since lending started
Banerjee et al.: The Miracle of Microfinance?
Vol.7 No. 1
39
Table 3—Self-Employment Activities: Revenues, Assets, and Profits (All households)
Investment
Assets in last 12
(stock) months Expenses
(1)
(2)
(3)
Panel A. Endline 1
Treated area
Observations
Control mean
Hochberg-corrected
p-value
Panel B. Endline 2
Treated area
Observations
Control mean
Hochberg-corrected
p-value
598
(384)
391*
(213)
6,800
2,498
6,800
280
1,261**
(530)
−134
(207)
6,142
5,003
6,142
1,007
255
(1,056)
Profit
(4)
Number
Has started Has closed a
Has a self-­
of selfa business in business in Index of
employment e­ mployment the last 12 the last 12 dependent
activity
activities
months
months
variables
(5)
(6)
(7)
(8)
(9)
6,685
4,055
354
(314)
6,239
745
0.0083
(0.0215)
0.018
(0.0380)
0.009
(0.006)
0.002
(0.008)
0.0357
(0.0188)
−530
(547)
542
(372)
0.023
(0.023)
0.045
(0.040)
−0.000
(0.010)
−0.000
(0.006)
0.0151
(0.0186)
6,116
5,225
6,090
953
6,810
0.349
6,142
0.418
6,810
0.503
6,142
0.561
6,757
0.047
6,142
0.083
2,352
0.037
6,142
0.053
6,810
0.000
0.175
6,142
0.000
>0.999
Notes: The table presents the coefficient of a “treatment” dummy in a regression of each variable on treatment (with control variables listed in the text). Cluster-robust standard errors in parentheses. Results are weighted to account for oversampling of Spandana
borrowers. The outcome variables are set to zero when the household does not have a business. Business outcomes are aggregated
at the household level when the households have more than one business. Information on closing a business in the year prior to the
endline 1 survey was only collected for those who had a business as of endline 1. Observations with missing or inconsistent itemized sales or revenues are dropped in columns 3 and 4. See online Appendix 1 for description of the construction of the profits, sales,
and inputs variables. All monetary amounts in 2007 Rs. Column 9 presents the coefficient of a “treatment” dummy in a regression
on treatment of an index of z-scores of the outcome variables in columns 1–8, plus revenues, number of new businesses, and number of new female-run businesses (see online Appendix Table A6, columns 1–3) for each round following Kling, Liebman, and Katz
(2007). p-values for this regression are reported using Hochberg’s step-up method to control the FWER across all index outcomes.
See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
later in the control group, however, households in the treatment group had, on average,
been borrowing for longer than those in the control group, which is reflected in the fact
that they had completed more loan cycles. On average, there was a difference of 0.085
loan cycles between the treatment and the control households at endline 2 (column 8),
which is almost unchanged from endline 1.14 The primary difference between treatment and control group at endline 2 is thus the length of access to microfinance. Since
microfinance loans grow with each cycle, treatment households also had larger loans.
Among those who borrow, there was by endline 2 a significant difference of about
Rs. 2,400 (or 14 percent) in the size of the loans (not reported). Since about one third
of households borrow, this translates into an (insignificant) difference of about Rs. 800
in average borrowing (column 3).
B. New Businesses and Business Outcomes
Panel A in Table 3 presents the results from the first endline on business outcomes. Column 7 indicates that the probability that a household starts a business
is in fact not significantly different in treatment and control areas. In comparison
14 This difference is no longer significant at EL2, possibly owing to recall error and to the fact that we only collected information on the maximum number of cycles borrowed from any MFI, so this figure does not distinguish,
e.g., a household that borrowed three cycles each from two lenders versus three cycles from one lender. 40
American Economic Journal: Applied Economics
January 2015
areas, 4.7 percent of households opened at least one business in the year prior to
the survey, compared to 5.6 percent in treated areas. However, treatment households were somewhat more likely to have opened more than one business in the past
year, and more new businesses were created in treatment areas overall: 6.8 per 100
households, versus 5.3 per 100 households in control areas.15 The 90 percent confidence interval on new business creation ranges from an additional 0.3pp to 2.6pp
additional new businesses. Overall, treatment households are no more likely to have
a business and do not have significantly more businesses (columns 5 and 6).
Consistent with the fact that Spandana lends only to women, and with the stated
goals of microfinance institutions, the marginal businesses tend to be female-operated. When we look at creation of businesses that are owned by women,16 we find
that almost all of the differential business creation in treatment areas is in female-operated businesses—there are 0.014 more female-owned businesses in treatment
households than in control households, an increase of 55 percent (see Table 7, column 8). Households in treated areas were no more likely to report closing a business, an event reported by 3.9 percent of households in treatment areas and 3.7
percent of the households in comparison areas (column 8).17
Treatment households invest more in durables for their businesses. Since only a third
of households have a business, and most businesses use no assets whatsoever, the point
estimate is small in absolute value (Rs. 391 over the last year, or a bit less than a third of
the increase in average MFI borrowing in treatment households), but the increment in
treatment is more than the total value of business durables purchased in the last year by
comparison households (Rs. 280), and is statistically significant (column 2).
The rest of the columns in the panel A of Table 3 report on current business status and last month’s input expenses and profits (exclusive of interest payments). In
these regressions, we assign a zero to those households that do not have a business,
so these results give us the overall impact of credit on business activities, including
both the extensive and intensive margins. Treatment households have more business
assets (although the t-statistic on the asset stock is only 1.56). The treatment effect
on expenses is positive but insignificant (column 3).18
Finally, there is an insignificant increase in business profits (column 4). Since
this data includes zeros for households that do not have a business, this answers the
question of whether microcredit, as it is often believed, increases poor households’
income by expanding their business opportunities. The point estimate, at Rs. 354 per
month, corresponds to a roughly 50 percent increase relative to the profits received
by the average comparison household. This is thus large in proportion to profits, but
it represents only a very small increase in disposable income for an average household—recall that the average total consumption of these households is about Rs. 6,500
15 See online Appendix Table A6, column 2 A business is classified as owned by a woman if the first person named in response to the question “Who is the
owner of this business?” is female. Only 74 out of 3,188 businesses have more than 1 owner. Classifying a business
as owned by a woman if any person named as the owner is female does not change the result. 17 It is possible that households not represented in our sample, such as households that had not lived in the area
for three years, may have been differentially likely to close businesses in treated areas. However, the relatively small
amount of new business creation makes general-equilibrium effects on existing businesses rather unlikely. 18 There is also a positive but insignificant effect on business revenues; see online Appendix Table A6, column 2. 16 Banerjee et al.: The Miracle of Microfinance?
Vol.7 No. 1
41
Table 3B—Self-Employment Activities: Revenues, Assets and Profits (Households with old businesses)
Panel A. Endline 1
Treated area
Observations
Control mean
Hochberg-corrected
p-value
Panel B. Endline 2
Treated area
Observations
Control mean
Hochberg-corrected
p-value
Assets
(stock)
(1)
Investment
in last 12
months
(2)
Revenue
(3)
Expenses
(4)
Profit
(5)
Employees
(6)
898
(1,063)
2,083
6,757
1,119
(698)
2,083
678
5,266
(3,720)
1,955
14,505
1,620
(3,257)
2,020
12,325
2,105*
(1,100)
1,624
2,038
−0.05
(0.0824)
2,088
0.41
1,682
(1,412)
1,878
10,301
−948
(588)
1,878
2,292
343
(1,263)
1,859
12,564
−2,644*
(1,491)
1,862
12,418
839
(945)
1,844
1,948
−0.12
(0.099)
1,878
0.46
Index of
dependent
variables
(7)
0.09
(0.0406)
2,088
0.00
0.057
−0.007
−0.0263
1,878
0.00
>0.999
Notes: The table presents the coefficient of a “treatment” dummy in a regression of each variable on treatment (with
control variables listed in the text). Cluster-robust standard errors in parentheses. Results are weighted to account
for oversampling of Spandana borrowers. The outcome variables are set to missing when the household does not
have an old business (i.e., one started more than a year prior to the survey). Business outcomes are aggregated at
the household level when households have more than one business. Observations with missing or inconsistent itemized sales or revenues are dropped in columns 3 to 5. See online Appendix 1 for description of the construction of
the profits, sales, and inputs variables. All monetary amounts in 2007 Rs. Column 7 presents the coefficient of a
“treatment” dummy in a regression on treatment of an index of z-scores of the outcome variables in columns 1–6 for
each round following Kling, Liebman, and Katz (2007). p-values for this regression are reported using Hochberg’s
step-up method to control the FWER across all index outcomes. See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
per month, and an increase of Rs. 354 per month in business revenues is certainly not
going to change the life of the average person who gets access to microcredit.
Looking at all businesses outcomes taken together, we find a 0.036 standard deviation increase in the standardized index of business outcomes, which is significant
with conventional standard errors but not once the multiple hypothesis testing across
different families of outcomes is taken into account ( p-value of 0.18).19
This is the ITT estimate, and part of the reason it is low is that few households took
advantage of microcredit in the treatment groups (and some did in the control as well).
The marginal borrower in the treatment group may also have fewer opportunities than
someone who was interested enough to borrow in the control group. This does not rule
out that the businesses of some specific groups could have benefited from the loan.
To look at this in more detail, we focus on businesses that were already in existence
before microcredit started. We do this in Table 3B.20 For businesses that existed before
Spandana expanded, we find an expansion in businesses (revenue, inputs, and investment). While most individual indicators are imprecise, the overall business index is
significant and positive, even after correcting for multiple inference (0.09 standard
19 It is significant even with this correction when we control for strata dummies. In Table 3, we show that households are no more or less likely to close a business in the last year; there is thus
no sample selection induced by microfinance. 20 42
American Economic Journal: Applied Economics
January 2015
OLS
9,000
Quantile treatment effect
90% C.I
6,000
3,000
0
−3,000
5
10
20
30
40
50
60
70
80
90 95
Percentile
Figure 3. Treatment Effect on Business Profits
(HHs who have an old business, endline 1)
Notes: Old businesses are businesses started at least one year before the survey. Confidence intervals are cluster-bootstrapped at the neighborhood level.
deviations, with a p-value of 0.057 after the correction). We find an average increase in
profits of Rs. 2,105 in treatment areas, which is statistically significant and represents
more than doubling, relative to the control mean of Rs. 2,038. This increase is not due
to a few outliers; however, it is worth nothing it is concentrated in the upper tail (quantiles 95 and above), as shown in Figure 3. At every other quantile, there is very little
difference between the profits of existing businesses in treatment and control areas.
There are 81 businesses above the ninety-fifth percentiles, far more than a handful,
but the ninety-fifth percentile of monthly profit of existing businesses is Rs. 15,050 (or
$1,640 at PPP), which makes them quite large and profitable businesses for this setting. The vast majority of the small businesses make very little profit to start with, and
microcredit does nothing to help them. This finding, that microcredit is most effective
in helping already profitable businesses, is contrary both to much of the rhetoric of
microcredit and to the view of microcredit skeptics.
Finally, we have seen that the treatment led to some more business creation, particularly the creation of female-owned businesses. In Figure 4, Table 3C and online
Appendix Table A4, we show more data on the characteristics of these new businesses. The quantile regressions in Figure 4 (profits for businesses that did not exist
at baseline) show that all new businesses between the thirty-fifth and sixty-fifth percentiles have significantly lower profits in treatment areas. Table 3C, column 5 shows
that the mean profit is not significantly different across treatment and control due to
the noisy data, but the median new business in treatment areas has Rs. 1,250 lower
profits, significant at the 5 percent level (not reported in tables, but shown in the figure). The average new business is also significantly less likely to have employees in
the treatment areas: the number of employees per new business is 0.29 in control and
only 0.11 in treatment (column 6). For new businesses, the index across all outcomes
Banerjee et al.: The Miracle of Microfinance?
Vol.7 No. 1
12,000
43
OLS
Quantile treatment effect
90% C.I
6,000
0
−6,000
−12,000
5
10
20
30
40
50
60
70
80
90 95
Percentile
Figure 4. Treatment Effect on Business Profits
(HHs who have new business, endline 1)
Notes: New businesses are businesses started less than one year before the survey. Confidence
intervals are cluster-bootstrapped at the neighborhood level.
Table 3C—Self-Employment Activities: Revenues, Assets, and Profits
(Households with new businesses, EL1 only)
Treated area
Observations
Control mean
Hochberg-corrected
p-value
Assets
(stock)
(1)
−873
(2,201)
356
8,411
Investment
in last 12
months
(2)
−706
(1,324)
356
2,418
Revenue
(3)
−8,167
(7,314)
332
17,423
Expenses
(4)
−5,013
(4,049)
339
12,114
Profit
(5)
−3,548
(3,813)
270
6,081
Index of
dependent
Employees variables
(6)
(7)
−0.195*
(0.112)
356
0.29
−0.0815
(0.0445)
356
0.00
0.280
Notes: The table presents the coefficient of a “treatment” dummy in a regression of each variable on treatment (with
control variables listed in the text). Cluster-robust standard errors in parentheses. Results are weighted to account
for oversampling of Spandana borrowers. The outcome variables are set to missing when the household does not
have a new business (i.e., one started less than a year prior to the EL1 survey). Business outcomes are aggregated at
the household level when the households have more than one business. Observations with missing or inconsistent
itemized sales or revenues are dropped in columns 3 to 5. See online Appendix 1 for description of the construction
of the profits, sales, and inputs variables. All monetary amounts in 2007 Rs. Column 7 presents the coefficient of a
“treatment” dummy in a regression on treatment of an index of z-scores of the outcome variables in columns 1–6
following Kling, Liebman, and Katz (2007). p-values for this regression are reported using Hochberg’s step-up
method to control the FWER across all index outcomes. See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
is negative (0.082 standard deviations) and significant with conventional levels but
not after correcting for multiple inference (corrected p-value: 0.28).
These results could in principle be a combination of a treatment effect and a selection effect, but since the effect on existing businesses suggests a treatment effect that
is close to zero for most businesses (and the point estimate is positive), the effect for
new businesses is likely due to selection—the marginal business that gets started in
44
American Economic Journal: Applied Economics
January 2015
treatment areas is less profitable than the marginal business in the control areas. The
hypothesis that the marginal business that gets started is different in the treatment
group gains some additional support in online Appendix Table 4, which shows a comparison of the industries of old businesses and new businesses, across treatment and
comparison areas.21 Industry is a proxy for the average scale and capital intensity of
a business, which is likely to be measured with less error than actual scale or asset
use. The industry composition of new businesses do differ. In particular, the fraction
of food businesses (tea/coffee stands, food vendors, kirana/small grocery stores, and
agriculture) is 8.5 percentage points (about 45 percent) higher among new businesses
in treatment areas than among new businesses in comparison areas, and the fraction
of rickshaw/driving businesses among new businesses in treatment areas is 5.4 (more
than 50 percent) percentage points lower. Both these differences are significant at the
10 percent level. Food businesses are the least capital-intensive businesses in these
areas, with assets worth an average of just Rs. 930 (mainly dosa tawas, pots and pans,
etc.). Rickshaw/driving businesses, which require renting or owning a vehicle, are the
most capital-intensive businesses, with assets worth an average of Rs. 12,697 (the bulk
of which is the cost of the vehicle).
Microcredit would be expected to lower the profitability threshold to start a business
if interest rates are lower than those of other sources of lending available to the households. Another explanation for both results could be that, due to the fact that Spandana
lends to women, the marginal businesses are more likely to be female-owned, and are
thus started in sectors in which women are active. Furthermore, businesses operated
by women generally tend to be less profitable, perhaps because of social constraints on
what women can do and how much effort they can devote to an enterprise.22
Panel B of Table 3 shows the results for the business performance variables at the
time of the second endline. As noted above, by this time treatment and control households are equally likely to have a microcredit loan, but loan amounts in treatment
areas are larger and borrowers have been borrowing for a longer time. The results
follow a clear pattern, consistent with the idea that control households now borrow at
the same rate. We find no significant difference in business creation in treatment and
control areas: the point estimate is virtually zero (the 90 percent confidence interval
ranges from 2pp fewer new businesses, to 2.5pp more). The new businesses are in the
same industries in treatment and control areas, and the negative effects for new businesses at the median have disappeared (results omitted). For the contemporaneous
flow investment outcomes such as new business creation, business assets acquired in
the previous year, etc. (columns 8 through 11) the point estimate is very close to zero
(however the standard errors are large). On the other hand, businesses in treatment
areas have significantly larger asset stocks (column 1), which reflects the cumulative effect of the past years during which they had a chance to borrow and expand.
Despite this, their profits are still not significantly larger, though the point estimate is
around 60 percent of the sample mean (with a t-statistic of around 1.5). As shown in
21 Respondents could classify their businesses into 22 different types, which we grouped into the following:
food, clothing/sewing, rickshaw/driving, repair/construction, crafts vendor, and “other.” 22 This is true in this data, and also found, for example, in Sri Lanka by de Mel, McKenzie, and Woodruff
(2009). Banerjee et al.: The Miracle of Microfinance?
Vol.7 No. 1
6,000
45
OLS
Quantile treatment effect
90% C.I
4,500
3,000
1,500
0
−1,500
5
10
20
30
40
50
60
70
80
90 95
Percentile
Figure 5. Treatment Effect on Business Profits
(Full sample of business owners, endline 2)
Figure 5, the positive increase is once again concentrated in the right tail, although
it starts being positive a little earlier, at the eighty-fifth percentile.
Overall, microfinance is indeed associated with (some) business creation: in
the first year, it does lead to an increase in the number of new businesses created,
particularly by women (though not in the number of households that start a business). However, these marginal businesses are even smaller and less profitable than
the average business in the area, the vast majority of which are already small and
­unprofitable. Microfinance does also lead to greater investment in existing businesses, and an improvement in the profits for the most profitable of those businesses.
For everyone else, business profits do not increase, and, on average, microfinance
does not help the businesses to grow in any significant way. Even after three years,
there is no increase in the number of employees of businesses that existed before
Spandana started its operation (Table 3B, column 6). Table 4 shows that total
self-employment income is unaffected by treatment.
C. Labor Supply
Access to credit can lead to an increase in labor supply to finance investment
or the purchase of durable goods which were out of reach before due to savings
and borrowing constraints. This is an area where different evaluations of microcredit have very different results, ranging from a worrying increase in labor supply
for teenagers in Augsburg et al. (2015) to steep decreases for everyone in Crépon
et al. (2015). Table 5 shows the impact of the program on labor supply. In endline 1, the household head and spouse in treatment households increase their overall
labor supply by an average of 3.18 hours (column 6; 90 percent CI: 0.84, 5.5). The
increase occurs entirely in the households’ own businesses (column 7), and there
is no increase in number of hours worked for wages (column 8): those hours may
46
American Economic Journal: Applied Economics
January 2015
Table 4—Income
Panel A. Endline 1
Treated area
Observations
Control mean
Hochberg-corrected p-value
Panel B. Endline 2
Treated area
Observations
Control mean
Hochberg-corrected p-value
Self employment
Index of dependent
(profit)
Daily labor/salaried
variables
(1)
(2)
(3)
354
(314)
6,239
745
−526
(358)
6,827
2,988
−0.0501
(0.0459)
6,832
0.000
>0.999
542
(372)
6,090
953
−141
(212)
6,142
5,514
0.0114
(0.0261)
6,142
0.000
>0.999
Notes: The table presents the coefficient of a “treatment” dummy in a regression of each variable
on treatment (with control variables listed in the text). Cluster-robust standard errors in parentheses. Results are weighted to account for oversampling of Spandana borrowers. Self-employment
income equals profit of a self-employment activity (summed across activities if multiple in the
household). Equal to zero for households with no self-employment activity. Daily labor/salaried income is income from employment other than self employment, summed across working
household members. See online Appendix 1 for description of the construction of the profit variable. Column 3 presents the coefficient of a “treatment” dummy in a regression on treatment of
an index of z-scores of the outcome variables in columns 1–2 for each round following Kling,
Liebman, and Katz (2007). p-values for this regression are reported using Hochberg’s step-up
method to control the FWER across all index outcomes. See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
be much less elastic, if the households do not fully choose them. However, we do
not find the increase in teenagers’ labor supply that is sometimes feared to be a
potential downside of microfinance and that Augsburg et al. (2015) find in Bosnia
(as adolescents are drawn into the business by their parents); indeed, teenage girls
work about two hours less per week in treatment than control areas (column 4),
and this difference is significant at the 5 percent level. There is no effect on teenage
boys’ hours (column 5). Given that there is an increase in work hours among adults
and a decrease among teens, the overall index is, not surprisingly, close to zero and
­insignificant. By endline 2, as control households have started borrowing, the difference between treatment and control disappears.
D. Consumption
Table 6 gives intent-to-treat estimates of the effect of microfinance on household
spending. Columns 1 and 3 of panel A show that there is no significant difference
between treatment and comparison households in total household expenditures—
either total or nondurable—per adult equivalent. The point estimate is essentially
zero in both cases and we can reject at the 5 percent level the null hypothesis that
there was a Rs. 85 per month increase in total consumption per adult equivalent
and a Rs. 57 per month increase in nondurable consumption (about 6 percent of the
Banerjee et al.: The Miracle of Microfinance?
Vol.7 No. 1
47
Table 5—Time Worked by Household Members
Hours worked over the past seven days, by age group:
All adults and teens
Teens
Household head and spouse
of which:
Total
(1)
Panel A. Endline 1
Treated area
Observations
Control mean
Hochberg-corrected
p-value
of which:
Self
Outside
employment activities
(2)
(3)
6,827
92.38
2.466
(2.361)
6,762
34.38
−2.033
(2.741)
−1.238
(1.544)
1.713
(2.162)
−2.951
(2.490)
0.739
(2.245)
6,762
58.01
Girls
(4)
Boys
(5)
−2.076** −0.026
(1.046)
(2.065)
Self
Outside
employment activities
(7)
(8)
Total
(6)
3.176**
(1.421)
2.710*
(1.474)
0.466
(1.418)
Index
of
­dependent
variables
(9)
0.00647
(0.0179)
2,174
7.94
1,866
25.12
6,827
57.79
6,827
25.83
6,827
31.96
6,849
0.000
>0.999
0.440
(0.948)
−1.387
(1.521)
0.991
(1.176)
1.703
(1.583)
−0.712
(1.488)
−0.00555
(0.0130)
Panel B. Endline 2
Treated area
Observations
Control mean
Hochberg-corrected
p-value
6,142
83.34
6,142
37.00
6,142
46.34
1,789
5.83
1,665
20.95
6,142
51.31
6,142
25.38
6,142
25.93
6,142
0.000
>0.999
Notes: Teens are household members aged 16 to 20. Adults are household members aged 21 and above. Total hours includes hours
worked in self-employment and in outside activities. It does not include time spent in housework. See online Appendix 1 for description of the construction of the self-employment variable. Column 9 presents the coefficient of a “treatment” dummy in a regression
on treatment of an index of z-scores of the outcome variables in columns 1–8 for each round following Kling, Liebman, and Katz
(2007). p-­­values for this regression are reported using Hochberg’s step-up method to control the FWER across all index outcomes.
See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
Table 6—Consumption (Per capita, per month)
Total
(1)
Panel A. Endline 1 Treated area
Observations
Control mean
Hochberg-corrected
p-value
Panel B. Endline 2
Treated area
Observations
Control mean
Hochberg-corrected
p-value
Durables Nondurable
(2)
(3)
Food
(4)
Health
(5)
Festivals
Home
Temptation
and
durable
Education
goods
celebrations good index
(6)
(7)
(8)
(9)
10.24
(37.22)
19.73*
(11.35)
−6.50
(31.81)
−12.11
(12.06)
−3.7
(11.51)
−2.061
(9.865)
−48.83
(51.53)
0.42
(9.88)
−45.45
(46.92)
−11.20
(17.88)
−22.54
(17.50)
12.16
(15.19)
6,827
1,419
>0.999
6,142
1,914
0.691
6,781
116
6,140
131
6,781
1,305
6,142
1,755
6,827
525
6,142
687
6,827
140
6,141
187
5,415
168
4,910
206
−8.785*
(4.92)
−14.16*
(8.09)
−0.051
(0.057)
−10.07
(6.61)
6.17
(4.12)
6,827
84
6,142
118
6,827
69
6,103
90
6,841
2.37
−0.0127
(0.0426)
6,142
2.66
Notes: Columns 1–8: Monthly per capita household expenditures. Temptation goods include alcohol, tobacco, betel leaves, gambling, and food consumed outside the home. Column 9 calculated on a list of 40 home durable goods (stock, not flow). Each asset
is given a weight using the coefficients of the first factor of a principal component analysis. The index, for a household i, is calculated as the weighted sum of standardized dummies equal to 1 if the household owns the durable good, 0 otherwise. See online
Appendix 1 for description of the construction of the consumption variables. p-values for the regression in column 1 (total consumption) reported using Hochberg’s step-up method to control the FWER across all outcomes. See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
48
American Economic Journal: Applied Economics
January 2015
average in control for consumption, and 4 percent for nondurable consumption)
increase.23 Hence, enhanced microcredit access does not appear to be associated
with any meaningful increase in consumption after 15 to 18 months. Of course, this
may partly be due to the fact that relatively few people borrow, and that some in the
control group borrow from another MFI.24
While there are no significant impacts on average consumption and nondurable
consumption, there are shifts in the composition of expenditure: column 2 shows
that households in treatment areas spent a statistically significant Rs. 19.73 more
per capita per month,25 or Rs. 237 per capita over the last year, on durables than did
households in comparison areas. Note that this is probably an underestimate of the
total effect of loans on durable purchases, since our measure would miss anyone
who borrowed more than a year before the survey (the survey was 15 to 18 months
after the centers opened) and immediately bought a durable with the loan. The most
commonly purchased durables include gold and silver, motorcycles, sarees (purchased in bulk, presumably mainly for weddings or as stock for a business), color
TVs, refrigerators, rickshaws, computers, and cellphones.
Columns 7 and 8 show that while there was no detectable change in nondurable spending otherwise, the increase in durable spending by treatment households
was essentially offset by reduced spending on “temptation goods” and festivals.
Temptation goods are goods that households in our baseline survey said that they
would like spend less on (this is thus the same list of goods for all households). They
include alcohol, tobacco, betel leaves, gambling, and food consumed outside the
home. Spending on temptation goods is reduced by about Rs. 9 per capita per month
(column 7). We also see in column 8 a large fall in festival spending per capita in
the previous year (Rs. 14 or 21 percent of the control level), both significant at the
10 percent level). Together, the average drop in consumption in temptation goods
and festivals is Rs. 23 per capita per month. The decrease in festival expenditures
does not come from large changes in large, very expensive ceremonies such as weddings (we see very few of them in the data) but rather appears to come from declines
at all levels of the distribution of spending on festivals.
The absolute magnitude of these changes is relatively small: for instance, the
Rs. 19.73 of increased durables spending per capita per month at endline 1 is approximately $2.14 at 2007 PPP exchange rates. However, this represents an increase of
about 17 percent relative to total spending on durable goods in comparison areas.
Furthermore, this figure averages over nonborrowers and borrowers, and would be
larger if it was attributed to borrowers alone.
Panel B of Table 6 reports on the impact effects at the time of the second endline,
when both treatment and control households have access to the microfinance program. The effects on both total per capita spending and total per capita nondurable
The 90 percent CIs are (−52, 72) for total consumption and (−59, 46) for nondurable consumption. For total consumption, the implied treatment on the treated (TOT) or IV estimate is a Rs. 122 (10.24/0.084),
or 9 percent, increase, and for nondurable consumption it is a Rs. 77 (6 percent) decrease. However, the 90 percent
confidence interval on the TOT estimate is wide, ranging from an increase of Rs. 857 (or 60 percent) to a decrease
of Rs. 613 (or 43 percent) in total consumption per capita. The width of the TOT confidence intervals stems, of
course, from the low first stage. 25 The 90 percent CI is (1, 39). 23 24 Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
49
spending (columns 1 and 3) are negative with t-statistics around 1. Spending on
temptation goods is still lower by about Rs. 10 per month (column 7), similar to endline 1, though the effect is now insignificant. The effect on festivals is now positive
but insignificant. There is also no difference on average in durable goods spending
in endline 2 (column 2). Given that the main difference between treatment and control households at endline 2 is that treatment households have been borrowing for
longer, this suggests that, in the second cycle, households in the treatment seem to
just repeat the first cycle with another durable (of roughly the same size), while
households in the control group also acquire a durable.
E. Microfinance as Social Revolution: Education, Child Labor,
and Women’s Empowerment?
The evidence so far suggests a different picture from the standard description of
the role of microfinance in the life of the poor: the pent-up demand for it is not overwhelming; many households use their loan to acquire a household durable, reducing
avoidable consumption to finance it; some invest in their businesses, but this does
not lead to significant growth in the profitability of most businesses. Another staple of the microfinance literature is that because the loans are given to women and
give them a chance to start their own businesses, this would lead to a more general
empowerment of women in the households, and this empowerment would in turn
translate into better outcomes for everyone in the household, including education,
health, etc. (e.g., CGAP 2009). Indeed, we see a significant increase in the number
of businesses managed by women in endline 1 (Table 7, column 9).26 To examine
whether this increase in women’s entrepreneurship translates into increased bargaining power for women, Table 7 examines the effects of access to microfinance on
measures of women’s decision making and children’s education and labor supply.
A finding of many studies of household decision making is that an increase in
women’s bargaining power leads to an increase in investments in children’s human
capital (see Thomas 1990 and Duflo 2003). However, we find that there is no change
in the probability that children or teenagers are enrolled in school (Table 7, columns 1, 2, 5, and 6), although we do see a reduction in teenage girls’ labor supply
(Table 5, column 5). There is no difference in spending on private school fees, or in
private versus public school enrollment (results omitted). There is also no difference
in the number of hours worked by girls or boys aged 5 to 15 (columns 3 and 4).
Because there are many possible proxies for womens empowerment and many
“social” outcomes we use the approach of Kling, Liebman, and Katz (2007) to
test the null hypothesis of no effect of microcredit on “social outcomes” against
the alternative that microcredit improves social outcomes. We construct an equally
weighted average of z-scores for 16 social outcomes; this method gives us maximal
power to detect an effect on social outcomes, if such an effect is present.27 Column 7
26 There is no difference in the number of women-run businesses between treatment and control in endline 2,
which is unsurprising since all areas have access to microfinance at that point. 27 The 16 outcomes we use are: indicators for women making decisions on each of food, clothing, health, home
purchase and repair, education, durable goods, gold and silver, investment; levels of spending on school tuition,
fees, and other education expenses; medical expenditure; teenage girls’ and teenage boys’ school enrollment; and
50
American Economic Journal: Applied Economics
January 2015
Table 7—Social Effects
Share of children
aged 5–15
in school
Hours worked
per child aged 5–15
over the past 7 days:
Share of teenagers
(aged 16–20)
in school
Girls
(1)
Panel A. Endline 1
Treated area
Observations
Control mean
Hochberg-corrected
p-value
Panel B. Endline 2
Treated area
Observations
Control mean
Hochberg-corrected
p-value
Boys
(2)
−0.016 −0.012
(0.013) (0.011)
3,035
0.919
3,073
0.918
0.015
(0.011)
0.007
(0.011)
2,755
0.923
2,746
0.928
Girls
(3)
Boys
(4)
−0.028
(0.202)
0.613
(0.743)
−0.037 −0.007
(0.024) (0.028)
0.092
(0.133)
−0.531*
(0.269)
0.021 −0.021
(0.024) (0.027)
3,035
0.594
2,755
0.286
3,073
0.577
2,746
1.379
Girls
(5)
2,174
0.338
1,789
0.329
Boys
(6)
1,866
0.429
1,665
0.474
Index of Number new
women’s self-employ.
indepenactivities
dence/ managed by Index of
empowerwomen dependent
ment
(all HHs) variables
(7)
(8)
(9)
0.007
(0.023)
6,862
−0.001
−0.011
(0.021)
6,142
−0.003
0.0143*** −0.008
(0.005)
(0.0097)
6,762
0.026
−0.005
(0.006)
6,142
0.047
6,862
0.000
>0.999
0.005
(0.009)
6,142
0.000
>0.999
Notes: In columns 1–4 the sample is restricted to households with children between the age of 5 and 15. In columns 5–6 the sample is restricted to households with teens between the age of 16 and 20. Column 7 is the effect on an equally weighted average
of z-scores for the 16 social outcomes: indicators for women making decisions on each of food, clothing, health, home purchase
and repair, education, durable goods, gold and silver, investment; levels of spending on school tuition, fees, and other education
expenses; medical expenditure; teenage girls’ and teenage boys’ school enrollment; and counts of female children under one year
and one- to two-years-old. Column 9 presents the coefficient of a “treatment” dummy in a regression on treatment of an index of
z-scores of the outcome variables in columns 1–9 for each round following Kling, Liebman, and Katz (2007). p-values for this
regression are reported using Hochberg’s step-up method to control the FWER across all index outcomes. See text for details.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
shows that there is no effect on the index of social outcomes (point estimate ​0.007​
standard deviations) and we can rule out an increase of more than one twentieth of
a standard deviation with 95 percent confidence.28
This suggests that there is no prima facie evidence that microcredit leads to important changes in household decision making or in social outcomes. Furthermore, this
null effect is not an artifact of observing households only in the very short run.
Nothing major changes by endline 2: the effect of microfinance access on the index
of women empowerment is still very small (indeed, slightly negative) and insignificant, and anything but a small effect can still be ruled out. Recall that we are comparing households who, by EL2, are equally likely to borrow: the main difference
by EL2 is that households in the treatment group have had greater access to microfinance for the first 18 months; this may limit power to detect differences in the social
outcomes at the community level.
counts of female children under one year and one- to two-years-old. We selected these outcomes because they
would likely be affected by changes in women’s bargaining power within the household. 28 The 95 percent CI is (−0.04, 0.05). The units are standard deviations. Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
51
IV. Conclusion
This study—the first and longest running evaluation of the standard group-lending loan product that has made microfinance known worldwide—yields a number of
results that may prompt a rethinking of the role of microfinance.
The first result is that, in contrast to the claims sometimes made by MFIs and others (including our partner), demand for microloans is far from universal. By the end
of our three-year study period, only 33 percent of households borrow from an MFI,29
and this is among households selected based on their relatively high propensity to
take up microcredit. This does not appear to be an anomaly: two other randomized
interventions that have a similar design (in Morocco and in Mexico) also find relatively low take-up, while another study in rural South India that focuses specifically
on take-up of microfinance also finds it to be low (Banerjee et al. 2013). Perhaps
despite evidence of high marginal rates of return among microbusinesses, e.g., de
Mel, McKenzie, and Woodruff (2008), most households either do not have a project
with a rate of return of at least 24 percent—the APR on a Spandana loan—or simply
prefer to borrow from friends, relatives, or money lenders due to the greater flexibility those sources provide, despite costs such as higher interest (from moneylenders)
or embarrassment (when borrowing from friends or relatives) (Collins et al. 2009).
For those who choose to borrow, while microcredit “succeeds” in leading some
of them to expand their businesses (or to start a female-owned business), it does not
appear to fuel an escape from poverty based on those small businesses. Monthly
consumption, a good indicator of overall welfare, does not increase for those who
had early access to microfinance, either in the short run (when we may have foreseen that it would not increase, or perhaps even expected it to decrease, as borrowers
finance the acquisition of household or business durable goods), or in the longer run,
after this crop of households have access to microcredit for a while and when those
in the former control group should be the ones tightening their belts. Business profits do not increase for the vast majority of businesses, although there are significant
increases in the upper tail of profitability. This study took place in a dynamic urban
environment, in a context of very high growth. Microcredit seems to have played
very little part in this growth, though it may have different impacts in other settings.
Furthermore, in the Hyderabadi context, we find that access to microcredit appears
to have no discernible effect on education, health, or women’s empowerment in the
short run. In the longer run (when borrowing rates are the same, but households in
the treatment groups have on average borrowed for longer), there is still no impact
on women’s empowerment or other social outcomes. The results differ from study
to study on these outcomes, but as a whole they don’t paint a picture of dramatic
changes in basic development outcomes for poor families.
Microcredit therefore may not be the “miracle” that it is sometimes claimed to
be, although it does allow some households to invest in their small businesses. One
reason may be that the average business run by this target group is tiny (almost none
of them have an employee), is not particularly profitable, and is difficult to expand
29 The take-up rate is 42 percent in treatment areas and 33 percent percent in control areas. 52
American Economic Journal: Applied Economics
January 2015
even in a high-growth context, given the skill sets of the entrepreneurs and their life
situations. And the marginal businesses that get created thanks to microcredit are
probably even less profitable and dynamic: we find that the average new business in
a microcredit treatment area is less likely to have an employee than the new business
in the control areas, and the median new business is even less profitable in treatment
versus control areas.
Nevertheless, microcredit does affect the structure of household consumption.
We see households invest in home durable goods and restrict their consumption of
temptation goods and expenditures on festivals and parties. They continue to do so
several years later, and this decrease is not due to a few particularly virtuous households, but seems to be spread across the sample. Similar declines in these types of
expenses are also found in all the other studies. Altered consumption thus does not
seem to be tied to the ideology of a particular MFI.
Microfinance affects labor supply choices as well: here we find that households
that have access to loans seem to work harder on their own businesses; in other settings, they are found to cut arduous labor elsewhere. Thus, microcredit plays its
role as a financial product in an environment where access to both credit and saving
opportunities is limited. It expands households’ abilities to make different intertemporal choices, including business investment. The only mistake that the microcredit
enthusiasts may have made is to overestimate the potential of businesses for the poor,
both as a source of revenue and as a means of empowerment for their female owners.
REFERENCES
Angelucci, Manuela, Dean Karlan, and Jonathan Zinman. 2015. “Microcredit Impacts: Evidence from
a Randomized Microcredit Program Placement Experiment by Compartamos Banco.” American
Economic Journal: Applied Economics 7 (1): 151–82.
Attanasio, Orazio, Britta Augsburg, Ralph De Haas, Emla Fitzsimons, and Heike Harmgart. 2015.
“The Impacts of Microfinance: Evidence from Joint-Liability Lending in Mongolia.” American
Economic Journal: Applied Economics 7 (1): 90–122.
Augsburg, Britta, Ralph De Haas, Heike Harmgart, and Costas Meghir. 2
015. “The Impacts of Microcredit: Evidence from Bosnia and Herzegovina.” American Economic Journal: Applied Economics
7 (1): 183–203.
Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan. 2
015. “The Miracle of
Microfinance? Evidence from a Randomized Evaluation: Dataset.” American Economic Journal:
Applied Economics. http://dx.doi.org/10.1257/app.20130533.
Banerjee, Abhijit, Arun G. Chandrasekhar, Esther Duflo, and Matthew O. Jackson. 2013. “The Diffusion of Microfinance.” Science 341 (6144).
Buera, Francisco J., Joseph P. Kaboski, and Yongseok Shin. 2011. “The Macroeconomics of Microfinance.” http://www.cfsp.org/sites/default/files/macroeconomics_of_microfinance.pdf.
Collins, Daryl, Jonathan Morduch, Stuart Rutherford, and Orlanda Ruthven. 2009. Portfolios of the
Poor: How the World’s Poor Live on Two Dollars a Day. Princeton: Princeton University Press.
Crépon, Bruno, Florencia Devoto, Esther Duflo, and William Parienté. 2
015. “Estimating the Impact
of Microcredit on Those Who Take It Up: Evidence from a Randomized Experiment in Morocco.”
American Economic Journal: Applied Economics 7 (1): 123–50.
de Mel, Suresh, David McKenzie, and Christopher Woodruff. 2008. “Returns to Capital in Microenterprises: Evidence from a Field Experiment.” Quarterly Journal of Economics 123 (4): 1329–72.
de Mel, Suresh, David McKenzie, and Christopher Woodruff. 2009. “Are Women More Credit Constrained? Experimental Evidence on Gender and Microenterprise Returns.” American Economic
Journal: Applied Economics 1 (3): 1–32.
DiNardo, John, Nicole M. Fortin, and Thomas Lemieux. 1996. “Labor Market Institutions and the
Distribution of Wages, 1973–1992: A Semiparametric Approach.” Econometrica 64 (5): 1001–44.
Vol.7 No. 1
Banerjee et al.: The Miracle of Microfinance?
53
Duflo, Esther. 2003. “Grandmothers and Granddaughters: Old Age Pensions and Intra-Household
Allocation in South Africa.” World Bank Economic Review 17 (1): 1–25.
Fulford, Scott. 2011. “The Effects of Financial Development in the Short and Long Run.” http://
fmwww.bc.edu/ec-p/wp741.pdf.
Giné, Xavier, and Robert M. Townsend. 2004. “Evaluation of Financial Liberalization: A General
Equilibrium Model with Constrained Occupation Choice.” Journal of Development Economics 74
(2): 269–307.
Heckman, James J. 1992. “Randomization and Social Program Evaluation.” In Evaluating Welfare and
Training Programs, edited by Charles F. Manski and Irwin Garfinkel, 201–30. Cambridge, MA:
Harvard University Press.
Hochberg, Yosef. 1988. “A Sharper Bonferroni Procedure for Multiple Tests of Significance.”
­Biometrika 75 (4): 800–802.
Kaboski, Joseph P., and Robert M. Townsend. 2011. “A Structural Evaluation of a Large-Scale
­Quasi-Experimental Microfinance Initiative.” Econometrica 79 (5): 1357–1406.
Karlan, Dean, and Jonathan Zinman. 2010. “Expanding Credit Access: Using Randomized Supply
Decisions To Estimate the Impacts.” Review of Financial Studies 23 (1): 433–64.
Karlan, Dean, and Jonathan Zinman. 2011. “Microcredit in Theory and Practice: Using Randomized
Credit Scoring for Impact Evaluation.” Science 332 (6035): 1278–84.
Kling, Jeffrey R., Jeffrey B. Liebman, and Lawrence F. Katz. 2007. “Experimental Analysis of Neighborhood Effects.” Econometrica 75 (1): 83–119.
Microcredit Summit Campaign. 2012. “State of the Campaign Report.” Washington, DC: Results Educational Fund.
MIX Market. 2009. “Profile for Spandana.” http://www.mixmarket.org/mfi/spandana.
Polgreen, Lydia, and Vikas Bajaj. 2010. “India Microcredit Faces Collapse From Defaults.”
New York Times, November 17. http://www.nytimes.com/2010/11/18/world/asia/18micro.
html?pagewanted=all&_r=0.
Tarozzi, Alessandro, Jaikishan Desai, and Kristin Johnson. 2
015. “The Impacts of Microcredit: Evidence from Ethiopia.” American Economic Journal: Applied Economics 7(1): 54–89.
Thomas, Duncan. 1990. “Intra-Household Resource Allocation: An Inferential Approach.” Journal of
Human Resources 25 (4): 635–64.
World Bank. 2005. “World Development Indicators.” http://data.worldbank.org/products/data-books/
WDI-2005.
World Bank. 2007. “World Development Indicators, Table 4.14: Exchange rates and prices.” http://
siteresources.worldbank.org/DATASTATISTICS/Resources/WDI07frontmatter.pdf.
This article has been cited by:
1. Orazio Attanasio, Britta Augsburg, Ralph De Haas, Emla Fitzsimons, Heike Harmgart. 2015.
The Impacts of Microfinance: Evidence from Joint-Liability Lending in Mongolia. American
Economic Journal: Applied Economics 7:1, 90-122. [Abstract] [View PDF article] [PDF with links]
2. Alessandro Tarozzi, Jaikishan Desai, Kristin Johnson. 2015. The Impacts of Microcredit:
Evidence from Ethiopia. American Economic Journal: Applied Economics 7:1, 54-89. [Abstract]
[View PDF article] [PDF with links]
3. Britta Augsburg, Ralph De Haas, Heike Harmgart, Costas Meghir. 2015. The Impacts of
Microcredit: Evidence from Bosnia and Herzegovina. American Economic Journal: Applied
Economics 7:1, 183-203. [Abstract] [View PDF article] [PDF with links]